Muhammad Kashif

Engineering notes / Muhammad Kashif

Claude Cowork and ChatGPT Work: Two Desks, One Rulebook

Claude Cowork and ChatGPT Work landed in my notes as two desks with the same temptation: let the agent live across laptop, phone, and chat apps until it feels like staff. Cross-device control, ninety-plus connectors, par…

Claude Cowork and ChatGPT Work: Two Desks, One Rulebook

Claude Cowork and ChatGPT Work: Two Desks, One Rulebook

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Claude Cowork and ChatGPT Work landed in my notes as two desks with the same temptation: let the agent live across laptop, phone, and chat apps until it feels like staff. Cross-device control, ninety-plus connectors, parallel fix-test-review agents — powerful, then dangerous. I kept both desks. I wrote one rulebook they both have to obey.

Two desks, different muscles

Cowork leans into computer use, folders, Slack, Gmail, Drive — the messy desktop where real work hides. ChatGPT Work leans into structured work surfaces, data agents, and the broader GPT ecosystem including Dots-shaped always-on helpers. I use Cowork when the artifact is files and apps on my machine. I use Work when the artifact is a report, table, or GPT-native workflow.

I do not run both with write access to the same production inbox. Dual write is how undo becomes archaeology.

Cross-device is a feature and a threat model

Laptop, phone, call; Slack, Teams, email on the roadmap — the notes described assigning tasks from a phone after pairing to a desktop Claude with folders connected. Convenience is real. So is the risk that a phone tap triggers a desktop agent with Drive scope while you are half-paying attention in a rickshaw.

My rule: phone can enqueue read-only research and draft requests. Phone cannot approve production mutates. Approvals wait until I am at a keyboard with the activity log open.

Human-in-the-loop is not a slogan

Rules plus an activity log are the minimum. I write the rules in a file the agent must read: which folders, which senders, which actions need a typed yes. The log gets reviewed on a cadence, not only after something feels weird.

Specialist Dots and parallel agents that fix, test, and review sound like a team. They get the same rule file. Parallelism without shared rules is three interns with root.

Connectors: ninety is not a virtue

Ninety-plus connectors across Slack, Notion, Jira, Docs look like completeness. I enable three for a pilot, measure, then expand. Every connector is an OAuth story and a data egress story.

Self-scheduling agents that book their own follow-ups stay on a dummy calendar until I trust the brief. My real calendar is not a playground.

Computer use from the phone, constrained

Install desktop, connect folders, pair phone, assign tasks — the flow works. I connect a scratch folder first, never the client vault. Slack and Gmail connections start read-only. Drive gets a dedicated pilot folder with nothing that would hurt to leak.

When the pilot earns trust, scopes widen slowly. Trust is earned by clean logs, not by a successful demo in front of a founder.

Image generation and memory on a leash

Image tools and long memory make the desks feel continuous. Continuous memory also means continuous retention of junk and secrets. I schedule memory cleanups and ban pasting credentials into either desk "just this once."

Once is how secrets become training residue in a product I do not control.

The shared rulebook

One classification scheme. One approval matrix. One activity-log format. One weekly review slot. Cowork and Work can disagree on UX. They cannot disagree on what "confidential" means or who can send email.

If a vendor feature fights the rulebook, the feature waits. I am not rewriting security because a button is pretty.

What I tell clients who want both

Pick a primary desk for the pilot. Mirror the rulebook. Expand connectors only after a clean two-week log. Budget for the human review time or admit you are buying theatre. Theatre does not get my production connectors.

Phone-enqueue etiquette

I only enqueue from my phone when I can complete a full sentence about the goal, the forbidden actions, and the deadline. Voice notes that say "handle my email" are rejected by policy — my own policy — because they are how Cowork earns too much trust on a bumpy road.

Work and Cowork both get that etiquette. The rulebook is one page and lives in both desks' project instructions.

Connector diet week

Every quarter I remove half the connectors from a desk for one week and note what actually hurts. Most "essential" connectors turn out to be comfort. Comfort connectors still carry OAuth risk. Diet weeks keep the rulebook honest.